Introduction: The Growing Need for Scalable Suicide Prevention

Suicide remains a pressing global public health issue, particularly among youth and marginalized populations. Traditional risk assessments often depend on self-reported symptoms and sporadic clinical evaluations, both of which can fail to identify individuals at imminent risk. In this context, the rise of digital health technologies—particularly artificial intelligence (AI)—offers a transformative opportunity to develop scalable, non-invasive methods for early suicide risk detection.

Among the most promising innovations is the use of AI to analyze web-based voice recordings, identifying subtle vocal biomarkers associated with suicidal ideation and behavior. Machine learning models have demonstrated the ability to distinguish individuals who died by suicide from controls with notable accuracy (area under the curve [AUC] up to 0.85) based on voice data collected up to 12 months prior to the event. Additionally, acoustic analysis of clinical and hotline recordings has shown promise in detecting high suicidality and predicting risk escalation, with classification accuracies ranging from 69% to 79%.^1,2

By combining standardized voice datasets with supervised learning algorithms grounded in established suicide theories, researchers are advancing the development of reliable, clinically relevant tools that can be integrated into primary care and telehealth workflows. Systematic reviews support the potential for AI-analyzed vocal markers to outperform traditional survey-based tools, especially in remote and virtual settings. This article explores the underlying science, emerging methodologies, and broader implications of AI-powered voice analysis in suicide risk prediction—marking a critical frontier in digital mental health.^3,4

Why Voice? The Science Behind Vocal Biomarkers of Suicide Risk

Mental health disorders such as depression and suicidal ideation produce physiological and neurological changes that often manifest in speech. This makes voice an especially compelling biomarker for detecting emotional and cognitive states associated with suicide risk. Acoustic features such as pitch variability, prosody, speech energy, pause frequency, and speech rate can reflect emotional distress, psychomotor retardation, and cognitive disorganization—hallmarks of suicidality.⁵

While these changes may be imperceptible to clinicians or caregivers, machine learning models can identify and interpret complex, multidimensional patterns within audio data. For instance, suicidal individuals often present with slower, flatter, and more breathy speech patterns, linked to alterations in spectral slope and formant bandwidths. AI models analyzing telehealth call recordings have achieved predictive accuracies as high as 99.85% in identifying high-risk speech segments. Other acoustic parameters—such as fundamental frequency, jitter, shimmer, and signal-to-noise ratio—also significantly differ between suicidal and non-suicidal individuals, including adolescents.

Advanced feature extraction techniques, such as Mel-Frequency Cepstral Coefficients (MFCCs), combined with deep learning architectures, have further enhanced the ability to discriminate suicidal ideation from depression and healthy controls. These findings underscore the potential of voice analysis to serve as a non-invasive, objective, and scalable method for identifying individuals at risk—providing a foundation for continuous mental health monitoring and early intervention in digital contexts.^6,7

Artificial Intelligence in Voice-Based Suicide Detection

AI plays a central role in unlocking the diagnostic potential of voice data for suicide risk prediction. Deep learning algorithms, including Convolutional Neural Networks (CNNs), are commonly used to analyze spectrograms—visual representations of audio frequencies over time—capturing complex spatial acoustic patterns. Recurrent Neural Networks (RNNs), long short-term memory (LSTM) models, and transformer-based architectures are employed to extract and interpret the temporal dynamics of speech.

Pretrained models such as wav2vec and Whisper, which leverage large-scale self-supervised learning, have further advanced the field by providing robust feature representations from raw audio inputs. These tools enhance model performance and generalizability across varied linguistic, cultural, and acoustic contexts.

Typically, AI models are trained on anonymized voice samples collected via web-based platforms, mobile apps, or chatbots. The outputs may take the form of binary classifications (e.g., at-risk vs. not at-risk) or continuous risk scores, enabling rapid and objective screening. These approaches offer several distinct advantages over traditional assessments: scalability to large populations, seamless integration into digital health platforms, and the potential for continuous remote monitoring.

Importantly, standardized datasets that include both vocal and linguistic elements—curated with attention to ethical, cultural, and clinical variables—are helping to build more transparent, explainable AI systems. These models not only predict suicide risk with impressive accuracy but also provide interpretable features that clinicians can use to inform care decisions. As these tools mature, they hold promise for deployment as adjunctive screening instruments in primary care, mental health clinics, emergency departments, and telehealth settings.^8,9

Reference:

1. Krautz AE, Volkening J, Raue J, Otte C, Eickhoff SB, Ahlers E, Langner J. Prediction of suicide using web based voice recordings analyzed by artificial intelligence. Sci Rep. 2025 Jul 4;15(1):23855. doi: 10.1038/s41598-025-08639-2. PMID: 40615574.

2. Min S, Shin D, Rhee SJ, Park CHK, Yang JH, Song Y, Kim MJ, Kim K, Cho WI, Kwon OC, Ahn YM, Lee H. Acoustic Analysis of Speech for Screening for Suicide Risk: Machine Learning Classifiers for Between- and Within-Person Evaluation of Suicidality. J Med Internet Res. 2023 Mar 23;25:e45456. doi: 10.2196/45456. PMID: 36951913; PMCID: PMC10131783.

3. Iyer R, Meyer D. Detection of Suicide Risk Using Vocal Characteristics: Systematic Review. JMIR Biomed Eng. 2022 Dec 22;7(2):e42386. doi: 10.2196/42386. PMID: 38875684; PMCID: PMC11041425.

4. Krautz AE, Volkening J, Raue J, Otte C, Eickhoff SB, Ahlers E, Langner J. Prediction of suicide using web based voice recordings analyzed by artificial intelligence. Sci Rep. 2025 Jul 4;15(1):23855. doi: 10.1038/s41598-025-08639-2. PMID: 40615574.

5. Iyer R, Nedeljkovic M, Meyer D. Using Vocal Characteristics To Classify Psychological Distress in Adult Helpline Callers: Retrospective Observational Study. JMIR Form Res. 2022 Dec 19;6(12):e42249. doi: 10.2196/42249. PMID: 36534456; PMCID: PMC9811648.

6. Yünden S, Ak M, Sert M, Gica S, Çinar O, Acar YA. Examination of Speech Analysis to Predict Suicidal Behavior in Depression. Eur Psychiatry. 2024 Aug 27;67(Suppl 1):S57–8. doi: 10.1192/j.eurpsy.2024.167. PMCID: PMC11860114.

7. Figueroa C, Guillén V, Huenupán F, Vallejos C, Henríquez E, Urrutia F, Sanhueza F, Alarcón E. Comparison of Acoustic Parameters of Voice and Speech According to Vowel Type and Suicidal Risk in Adolescents. J Voice. 2024 Aug 30:S0892-1997(24)00254-6. doi: 10.1016/j.jvoice.2024.08.006. Epub ahead of print. PMID: 39217086.

8. Parsapoor Mah Parsa M, Koudys JW, Ruocco AC. Suicide risk detection using artificial intelligence: the promise of creating a benchmark dataset for research on the detection of suicide risk. Front Psychiatry. 2023 Jul 24;14:1186569. doi: 10.3389/fpsyt.2023.1186569. PMID: 37564247; PMCID: PMC10411603.